---

I. 3 Key Updates

Key Update 1｜Alibaba Releases Qwen3.8-Max — "QwenWork" Public Beta Launches Simultaneously

On August 3, Alibaba officially released the strongest model in the Qwen family to date, Qwen3.8-Max: built on the Qwen3.5 architecture, with 2.4 trillion parameters and 95B active parameters, supporting a 1 million token context window. Capabilities in coding, office tasks, scientific research, and long-horizon tasks have been comprehensively improved, enabling thousands of rounds of long-range interaction. Next week, the Qwen3.8-Max model weights will be open-sourced, along with Qwen3.8-27B—this is the first open-source flagship model in the Qwen series. API pricing: Domestic input 12 RMB, output 36 RMB/million tokens. On the same day, the enterprise-level Agent product "QwenWork" launched its public beta: integrating three products—Qoderwork, Mulerun, and Wukong—it has initially connected with DingTalk. The web version and PC client are open for experience; an independent App and international version will follow later.

LLM Expert｜ Activating only 95B out of 2.4T total parameters means taking the route of "large model capabilities, small model bills" to the end—according to official pricing, input costs are an order of magnitude lower than comparable overseas flagships. More noteworthy is the "first open-source flagship": previously, Qwen's strongest models were always closed-source. Once the weights are released next week, developers will for the first time have access to a self-deployable Chinese model at the flagship tier. What truly remains to be verified is the stability of long-horizon Agents; "thousands of interactions without decay" still needs third-party replication.

Software Expert｜ Ordinary users can take action today: apply for the public beta on the "QwenWork" website and prioritize testing two types of tasks—long document batch processing (contracts, weekly reports, bids) and daily approval workflows in DingTalk. For teams already using DingTalk, migration costs are nearly zero; but during the beta period, run non-sensitive tasks for a week to check stability first, and don't migrate core production data all at once. Teams with development capabilities can wait for the open-source weights next week before evaluating private deployment.

---

Key Update 2｜Grok Adds Video Understanding — Musk Tests with a Fake Kobe Bryant Video

On August 2 local time, Grok officially launched video interaction features: users can upload local videos or paste video links in the chat box just like images, allowing the model to summarize the entire content, identify characters and objects in the scene, break down action details, and ask continuous follow-up questions about any detail. Compared to the old version which only extracted audio text to generate summaries, the new version performs full-dimensional understanding by linking consecutive frames, audio, and context. Musk demonstrated on X platform: uploading an AI-generated video featuring the late basketball star Kobe Bryant as the prototype, Grok analyzed the frames, extracted dialogue, and combined background facts to conclusively determine the video was Deepfake content.

LLM Expert｜ Every major player has video understanding; the value here lies in "judgment" rather than "description": the model cross-referenced visuals, dialogue, and the background fact that "Kobe passed away in 2020" to reach the Deepfake conclusion—this signals multimodal AI moving from "seeing" to "thinking." But one demo doesn't equal stable capability; official demos usually cherry-pick samples, so real detection rates await third-party benchmarks.

Software Expert｜ Three scenarios ordinary users can use today: automatically generating meeting minutes from screen recordings, breaking online course videos into chapters with summaries, and verifying suspicious videos when scrolling. Suggest adding "upload video and ask three questions" directly into your daily workflow. A reminder: fake video detection should only serve as auxiliary reference; for major judgments, don't trust just one model—cross-checking with two tools is more reliable.

---

Key Update 3｜German Court Rules Suno Infringing — End of the Era of "Free-Riding" Training for AI Music

The Munich Regional Court ruled on July 31 in favor of the German music copyright society GEMA against AI music generation platform Suno, a case that continued to ferment over the weekend. The court found that Suno used GEMA's song library to train its model and stored/copied relevant works, violating German and US copyright laws, rejecting its fair use defense, covering both model training and output generation; Suno must cease infringing usage, disclose revenue information, and pay damages (amount TBD), with the option to appeal. The six songs involved include "Rasputin," "Forever Young," and "Mambo No. 5." This is Europe's first ruling that "the model remembering a song" itself constitutes copying.

LLM Expert｜ The technical implication of the verdict is more important than the compensation amount: the court accepted the argument that "reproducible protected expression within the model constitutes copying," effectively negating the classic defense that "models only store mathematical parameters, not works." If this logic spreads, compliance for training data in all generative models (text, image, etc.) will need re-evaluation, similar lawsuits will emerge in the European market, and the window for "train first, pay later" is closing.

Software Expert｜ For users creating content with AI, three practical suggestions: First, for commercial projects (ads, short video BGM), prioritize tools with published copyright authorization chains; Second, don't treat paid subscriptions as a copyright shield—they only grant usage rights, not guaranteeing generated outputs are non-infringing; Third, if you're using Suno for commercial orders, audit existing projects and wait for the appeal outcome to become clear before scaling up major commercial uses. Personal entertainment use is not directly affected for now.

---

II. 10 Selected Updates

1. DeepSeek-V4-Flash Official Version Lands on National Supercomputing Internet (Aug 2): V4-Flash-0731, after extensive post-training, sees significantly enhanced agent capabilities and instruction following, with performance on multiple benchmarks rivaling current top closed-source models; no environment configuration needed for one-click API calls, and model files can be downloaded from the community to complete fine-tuning, deployment, and adaptation in a trusted Notebook environment. → No hassle with GPUs and environments; access flagship-tier open-source models via a webpage, lowering the barrier for SMEs even further.

2. Claude Opus 4.1 Retires Tomorrow Morning (Aug 5): Anthropic officially requires apps still calling claude-opus-4-1-20250805 to migrate to claude-opus-4-8; requests on their own platform will fail after retirement; timelines for Amazon Bedrock and Google Cloud need separate confirmation. → Today is the last day to check model IDs in your API code; after changes, ensure regression testing for prompts, tool calls, and costs.

3. OpenRouter Latest Weekly Chart: Top 5 Global Call Volume All Chinese LLMs (Aug 2): Xiaomi MiMo-V2.5 tops the list, with single-week token call volume reaching 10.46T, the only model exceeding 10T that week, growing approx. 616% since May; DeepSeek, Tencent Hunyuan, etc., also rank high. → The "default model" underlying the overseas tools you use is increasingly likely to be a Chinese open-source model; remember to add domestic models to your candidate list when comparing tool options.

4. Tencent CodeBuddy Domestic Version Fully Supports DeepSeek-V4-Flash Official Version (Aug 4): Model lists updated simultaneously across IDE, plugin, and CLI; officially stated significant improvement in Agent capabilities; the same list includes Hy3, GLM-5.1/5.2/5v, Kimi-K2.6/2.7/3, MiniMax-M3, DeepSeek-V4 Pro, etc. → Developers using CodeBuddy can switch their main model to the V4-Flash official version; Agent task usability is clearly better than during the preview period.

5. Claude Cowork Expands to Mobile and Web: Sessions and files sync across devices, supporting background work, scheduled tasks, shared chats and projects, mobile approvals; Max users first, gradually opening up, double usage quota limited-time until Aug 5. → You can assign work to Claude while traveling; if you want the double quota, schedule heavy tasks before tomorrow night.

6. Tencent Yuanbao Upgrades Image Understanding, Promoting "Snap When Stuck" (Aug 3): Take a photo to recognize complex scenes, extract key data, and recommend next-step automation commands, covering all life scenarios. → Usable even if you can't write prompts—snap then ask, this is the most beginner-friendly version of Yuanbao for elders and novices.

7. SenseNova Open-Sources SenseNova U1.5-Lite-Preview: 8B Scale Native 4K Image Generation (Aug 3): Based on NEO-Unify architecture, a single 8B-MoT model integrates visual understanding, reasoning, generation, and editing, surpassing U1 in multiple mainstream evaluations and matching commercial closed-source models; code and weights available on GitHub, Hugging Face, and ModelScope. → Local deployment of 4K image generation enters the 8B scale for the first time, runnable on consumer-grade GPUs; those making posters and infographics can download and try directly.

8. Google Pauses Google Earth AI Image Generation Feature (Aug 3): The feature was paused less than 48 hours after launch, officially to strengthen security safeguards, restoration time undetermined. → The rhythm of "launch first, secure later" for generative features is tightening; if you have workflows relying on this feature, keep backup tools ready.

9. Huawei Noah Open-Sources MindMemOS: Giving Agents "Portable Memory" (Aug 2): A transferable, self-evolving memory operation layer for AI Agents, open-sourced under MIT license, offering API, SDK, CLI, and plugin integration methods; cloud service registration and trial are open. → Developers building Agent applications don't need to reinvent the memory module for every project; ordinary users will naturally perceive "AI remembers you" once upper-layer apps integrate it.

10. Lingguang App Flash App Creators Surpass 4 Million (Aug 3): Officially stated the vast majority are ordinary users without programming backgrounds, indicating ordinary people are becoming the main force in AI app creation. → "Making apps without knowing code" is already a scaled reality; if you want to make your own small tools, starting by imitating popular flash apps is the fastest way.

---

III. News Briefs

• Moonshot AI responds to media rumors: Reports of "submitting HK IPO application as early as this month" are false (AIBase, Aug 4).
• DeepSeek: V4-Pro official version to be released "soon", V4-Flash official version API public beta ongoing (IT Home, Aug 2).
• QwenWork will launch an independent App and international version later, achieving multi-end linkage (Jiupai Finance / Yicai, Aug 3).
• Qwen3.8 API Pricing: Domestic input 12 RMB, output 36 RMB/million tokens, implicit cache hit 1.5 RMB; Overseas input